In a recent article published in natural medicineResearchers apply artificial intelligence (AI) techniques to long-term, real-world clinical data in a surveillance program for early detection of patients at high risk for pancreatic cancer, one of the most aggressive diseases designed.
Background
The incidence of pancreatic cancer is increasing and is the leading cause of cancer-related death worldwide. Diagnosing pancreatic cancer is difficult due to the lack of understanding of pancreatic cancer risk factors.
Delayed detection in advanced or distant metastatic disease hampers treatment and makes patient survival very rare. Only 2-9 percent of such patients survive after 5 years.
Although age is a recognized risk factor for pancreatic cancer, population-wide screening based on age is impractical due to the high cost of laboratory testing and the possibility of false-positive results.
Additionally, family history data and genetic risk factors for the general population are often unavailable. Therefore, there is an urgent need to develop affordable surveillance programs for early detection of pancreatic cancer in the general population.
About research
In the current study, researchers used actual long-term clinical records of a large number of patients to identify a significant number of patients at high risk for pancreatic cancer.
They exploit recently developed machine learning (ML) techniques using patient records from the Danish National Patient Registry (DNPR) and later the US Department of Veterans Affairs (US-VA) Corporate Data Warehouse (CDW). bottom.
The former includes data on 8.6 million patients collected between 1977 and 2018, representing 24,000 pancreatic cancer cases, while the latter includes pancreatic cancer cases. It included clinical data from 3 million patients, including 3,900 cases.
The research team trained and tested different ML models based on a set of disease codes in DNPR and US-VA clinical records and tested their prediction of cancer development within incremental time intervals called CancerRiskNet.
In building the predictive model, the research team used three-letter categories of the International Classification of Diseases (ICD) diagnostic codes, and defined patients with at least one code below C25, indicating malignant neoplasms of the pancreas, as ‘pancreatic cancer patients’. defined as
The cancer diagnostic disease code accuracy was approximately 98%. Finally, the researchers flagged which diagnoses were most informative for cancer risk in the patient’s diagnostic code history in order to suggest an ideal surveillance program.
In addition, researchers evaluated the predictive performance of different models trained on DNPR using receiver operating characteristic area under (AUROC) curves and relative risk (RR) curves. In addition, we reported ML-derived RR scores for cancer patients in the high-risk group.
result
All previous studies that used real-world clinical records to predict pancreatic cancer risk have yielded promising results, but the time series of medical history was used to extract longitudinal features of the time series. I didn’t use it. In this study, we evaluated a non-time series model for the DNPR dataset.
Overall, the time series model Transformer performed best in predicting cancer incidence within 36 months from the date of assessment, followed closely by GRU with AUROC of 0.879 and AUROC of 0.852.
The model had a RR of 104.7 at operational time defined by n = 1,000 of the highest risk patients out of 1 million patients.
The performance of the bag-of-words and MLP models for predicting cancer incidence within 36 months for AUROC was 0.807 and 0.845, respectively. However, compared to Transformer, bag-of-words and MLP had much lower RRs (104.7 vs. 2.1 and 26.6).
Excluding data, i.e., disease diagnoses entered in the past 3, 6, and 12 months prior to pancreatic cancer diagnosis, from AUROC with a best model performance of 0.879, 3/6 dropped to AUROCs of 0.843, 0.829, and 0.827. /12 months.
This analysis showed that the ML model trained on data from both sources had a positive predictive value (PPV) of 0.32 over a 12-month prediction period. That means about 320 patients will eventually develop pancreatic cancer.
Physicians may have identified some cases based on recognized risk factors for pancreatic cancer, such as chronic pancreatitis, but a conservative approximation suggests that some of those, nearly 70, are new. will be specified.
Despite using common ICD disease codes and similar cancer survival rates, cross-application of DNPR data to US-VA data results in poorer ML model performance and regionally optimal model performance increased need for independent model training across geographic regions to .
However, the ideal scenario for multiple institutions working together to achieve a globally relevant set of predictive rules would require federated learning across different healthcare systems.
Conclusion
The predictive accuracy of the ML-based models described in this study depends on data beyond disease codes (e.g., observations written in clinical notebooks, test results, genetic profiles of more people, readings from wearable devices). health-related information) may be improved.
Therefore, clinical implementation of early diagnosis of pancreatic cancer requires identification of high-risk patients.
Because the patients at highest risk are a smaller subset of the large computer-screened population, expensive and sophisticated clinical screening and intervention programs are limited to a small number of patients. Become.
Nevertheless, the introduction of AI into real-world clinical records has the potential to shift the focus from late-stage to early-stage cancer care, resulting in significantly improved quality of life for all patients, The cost-benefit ratio will also improve. About cancer treatment.
Quote
To cite this article in your essay, paper or report, please use one of the following formats:
-
APA
Mathur, Neha. (May 10, 2023). Could artificial intelligence help predict pancreatic cancer risk using disease trajectories? News Medical. From https://www.news-medical.net/news/20230510/Can-artificial-intelligence-help-predict-the-risk-of-pancreatic-cancer-using-disease-trajectories.aspx He May 2023 Obtained on the 10th.
-
MLA
Mathur, Neha. “Could artificial intelligence help predict pancreatic cancer risk using disease trajectory?” News – Medical. May 10, 2023.
-
Chicago
Mathur, Neha. “Could artificial intelligence help predict pancreatic cancer risk using disease trajectory?” News Medical. https://www.news-medical.net/news/20230510/Can-artificial-intelligence-help-predict-the-risk-of-pancreatic-cancer-using-disease-trajectories.aspx. (Accessed May 10, 2023).
-
Harvard University
Mathur, Neha. 2023. Could artificial intelligence help predict pancreatic cancer risk using disease trajectory?. News-Medical, accessed May 10, 2023, https://www.news-medical.net/news/20230510/Can-artificial-intelligence-help-predict-the-risk-of-pancreatic-cancer-using- disease-trajectories.aspx.
